Before pulling a NIM image, run the pre-download profile preflight for your exact GPU and free VRAM, and choose the profile it reports rather than the default. Budget real host RAM well above the admission floor, and keep the failed container around to read its logs when a cold start dies. Free unrelated GPU memory before retrying instead of hammering the same launch.

Context: NVIDIA's Cosmos3 NIM operations guide documents the profile compatibility traps. 'No compatible profile' means the GPU's compute capability, per-device VRAM, or system memory does not satisfy the selected model, precision, and offload requirements, so the fix is to run the pre-download profile preflight and pick a compatible profile. Separately, a passing preflight does not guarantee a successful cold start: the admission floor is only 16 GiB, and practical host RAM or artifact materialization can still fail the start.